CN113869521A

CN113869521A - Method, device, computing equipment and storage medium for constructing prediction model

Info

Publication number: CN113869521A
Application number: CN202010612047.9A
Authority: CN
Inventors: 张彦芳; 孙旭东; 常庆龙; 张亮
Original assignee: Huawei Technologies Co Ltd
Current assignee: Huawei Technologies Co Ltd
Priority date: 2020-06-30
Filing date: 2020-06-30
Publication date: 2021-12-31
Also published as: US20230146912A1; EP4167149A1; WO2022001918A1; EP4167149A4

Abstract

The application provides a method, a device, a computing device and a storage medium for constructing a prediction model, and belongs to the technical field of artificial intelligence. The method comprises the following steps: the method comprises the steps of obtaining a model search space corresponding to a target prediction scene according to a target data set and/or scene information of the target prediction scene, wherein the model search space comprises a model and a hyper-parameter, carrying out model training according to the model and the hyper-parameter included in the target data set and the model search space to obtain a prediction model which is trained, and obtaining a prediction model corresponding to the target prediction scene according to an evaluation result of the prediction model which is trained. By the method and the device, the efficiency of constructing the prediction model can be improved.

Description

Method, device, computing equipment and storage medium for constructing prediction model

Technical Field

The present application relates to the field of artificial intelligence technologies, and in particular, to a method, an apparatus, a computing device, and a storage medium for constructing a prediction model.

Background

With the development of computing technology, prediction models are generally used to predict some phenomena occurring in the future, so that the prediction models need to be constructed.

In the related technology, when a prediction model is constructed, a model search space comprises all collected machine learning models at present, one machine learning model is randomly selected from the model search space, the machine learning model is subjected to hyper-parameter optimization and training based on sample data, finally, the performance of the trained prediction model is evaluated, and whether all the machine learning models in the model search space are tried or not is judged. If not, selecting one machine learning model in the model search space again for training until all the machine learning models in the model search space are tried, and selecting the prediction model with the best performance evaluation result as output; if all the tests are tried, the prediction model with the best performance evaluation result is selected as output.

Because the number of machine learning models included in the model search space is large, all the machine learning models are trained, a large amount of processing resources are consumed to obtain the prediction model, and the efficiency of determining the prediction model is low.

Disclosure of Invention

The application provides a method, a device, a computing device and a storage medium for constructing a prediction model, and the efficiency of determining the prediction model can be improved.

In a first aspect, a method for constructing a prediction model is provided, the method comprising: the method comprises the steps of obtaining a model search space corresponding to a target prediction scene according to a target data set and/or scene information of the target prediction scene, wherein the model search space comprises a model and a hyper-parameter, carrying out model training according to the model and the hyper-parameter included in the target data set and the model search space to obtain a prediction model which is trained, and obtaining a prediction model corresponding to the target prediction scene according to an evaluation result of the prediction model which is trained.

In the solution shown in the present application, in the first aspect, the method of constructing the prediction model is performed by a construction apparatus. The construction equipment can obtain a model search space corresponding to the target prediction scene by using a target data set of the target prediction scene and/or scene information of the target prediction scene, and then obtain a prediction model which is trained by using a model and hyper-parameters included in the target data set and the model search space. And the construction equipment evaluates each prediction model after training to obtain the evaluation result of each prediction model. And the construction equipment determines a prediction model corresponding to the target prediction scene based on the evaluation result of each prediction model. Therefore, the construction equipment can acquire the model search space of the target prediction scene, so that the number of selectable models in the model search space is relatively small, all models and hyper-parameters in the model search space can be quickly combined to complete training, the construction time of the prediction model can be further saved, and the efficiency of determining the prediction model is improved.

In a possible implementation manner, obtaining a model search space corresponding to a target prediction scene according to a target data set of the target prediction scene and/or scene information of the target prediction scene includes: acquiring an initial model search space, wherein the initial model search space is obtained by analyzing a historical data set, the historical data set is a target data set of a target prediction scene, or data in the historical data set is similar to data in the target data set; screening models and/or hyper-parameters in an initial model search space according to a target data set and/or scene information of a target prediction scene to obtain a model search space corresponding to the target prediction scene, wherein the models in the model search space are partial models or all models in the initial model search space, and the hyper-parameters in the model search space are partial hyper-parameters or all hyper-parameters in the initial model search space; or acquiring a model search space corresponding to the target prediction scene from a preset model and the hyper-parameters according to the target data set of the target prediction scene and/or the scene information of the target prediction scene.

According to the scheme, the construction equipment can obtain an initial model search space obtained by using the target data set, then the construction equipment uses the target data set and/or scene information of the target prediction scene, in the initial model search space, the model and/or the hyper-parameter are screened, and the screened model and/or the hyper-parameter form a model search space corresponding to the target prediction scene. Or the construction equipment stores preset models and hyper-parameters, the construction equipment uses the target data set and/or scene information of the target prediction scene, the models and the hyper-parameters are screened from the preset models and the hyper-parameters, and the screened models and the hyper-parameters form a model search space corresponding to the target prediction scene. Therefore, only the model and the hyper-parameter related to the prediction scene are screened out, so that the number of selectable models in the model search space is relatively small, all the models and the hyper-parameter in the model search space can be quickly combined to complete training, the construction time of the prediction model can be further saved, and the efficiency of determining the prediction model is improved.

In one possible implementation manner, the model in the model search space is a model in which the features of the data set applicable in the initial model search space have similarity with the first features of the target data set, and the hyper-parameters in the model search space are hyper-parameters in which the features of the data set applicable in the initial model search space have similarity with the first features of the target data set; or the model in the model search space is a model with similarity between the scene information applicable to the initial model search space and the scene information of the target prediction scene, and the hyper-parameter in the model search space is a hyper-parameter with similarity between the scene information applicable to the initial model search space and the scene information of the target prediction scene; or the model in the model search space is a model in which the scene information applied in the initial model search space has similarity with the scene information of the target prediction scene and the characteristics of the applied data set have similarity with the first characteristics of the target data set, and the hyper-parameters in the model search space are hyper-parameters in which the scene information applied in the initial model search space has similarity with the scene information of the target prediction scene and the characteristics of the applied data set have similarity with the first characteristics of the target data set.

In the solution shown in the present application, the first feature of the target data set may be a partial feature or a full feature of the target data set. The construction equipment screens out the model and the hyper-parameters in the initial model search space by using the target data set and/or the scene information of the target prediction scene, and the screened model and the hyper-parameters form a model search space corresponding to the target prediction scene. In this way, only models and hyper-parameters relevant to the predicted scenario may be filtered out.

In one possible implementation, obtaining an initial model search space includes: determining a model and a hyper-parameter corresponding to a target prediction scene according to a historical data set, and forming an initial model search space by the determined model and the hyper-parameter; or sending a model acquisition request to the cloud device, wherein the model acquisition request is used for requesting to acquire an initial model search space; and receiving an initial model search space sent by the cloud equipment.

According to the scheme, the construction equipment can determine the initial model search space by the construction equipment, and can also request the initial model search space from the cloud equipment. Thus, the initial model search space can be acquired more flexibly.

In one possible implementation, determining a model and a hyper-parameter corresponding to a target prediction scenario according to a historical data set includes: and obtaining a model and a hyper-parameter with the similarity between the characteristics of the applicable data set and the second characteristics of the target data set, and determining the model and the hyper-parameter in the initial model search space corresponding to the target prediction scene.

According to the scheme, the construction equipment can acquire the model and the hyper-parameters with the similarity between the characteristics of the applicable data set and the second characteristics of the target data set in the preset model and the hyper-parameters. And the construction equipment forms an initial model search space by using the obtained model and the hyper-parameters. In this way, partial models and hyper-parameters may be first filtered out through the target dataset.

In one possible implementation, the second feature of the target dataset is the same feature or a different feature than the first feature of the target dataset.

According to the scheme, under the condition that an initial model search space exists and a target data set is not used for screening the model and/or the hyper-parameters in the initial model search space, the second characteristic of the target data set and the first characteristic of the target data set can be the same characteristic or different characteristics. In the case where an initial model search space exists and the model and/or hyper-parameters in the initial model search space are filtered using the target dataset, the second feature of the target dataset is a different feature than the first feature of the target dataset. In this way, the preset model and the hyper-parameters can be screened by using the characteristics of more target data sets, so that the model and the hyper-parameters in the model search space corresponding to the target prediction scene can be more consistent with the target prediction scene.

In a possible implementation manner, the scene information of the target prediction scene includes a calculation performance requirement and/or a prediction requirement, the calculation performance requirement includes one or more of memory information, Central Processing Unit (CPU) information, or inference speed, and the prediction requirement includes a prediction duration and/or a prediction period for predicting by a prediction model corresponding to the target prediction scene. Therefore, the prediction model corresponding to the target prediction scene can better meet the requirement.

In one possible implementation, the first feature of the target data set includes classification information and/or statistical information, wherein the classification information includes one or more of period information, fluctuation information or mutation information of the time series in the target data set, and the statistical information includes one or more of sampling interval, sampling duration or missing acquisition condition of the time series in the target data set. Therefore, the prediction model corresponding to the target prediction scene can better meet the requirement.

In a possible implementation manner, obtaining a prediction model corresponding to a target prediction scene according to an evaluation result of a prediction model that has been trained includes: and selecting a prediction model with the optimal evaluation result from the trained prediction models, and determining the prediction model as a prediction model corresponding to the target prediction scene. In this way, a prediction model with the best performance can be selected for the target prediction scene.

In a possible implementation manner, before model training is performed according to a model and a hyper-parameter included in a target data set and a model search space to obtain a trained prediction model, the method further includes: obtaining a search strategy corresponding to a target prediction scene, wherein the search strategy comprises a model search strategy; according to the model and the hyper-parameters included in the target data set and the model search space, model training is carried out to obtain a prediction model which is trained, and the method comprises the following steps: searching the model and the hyper-parameters in a model search space according to a model search strategy; and performing model training according to the target data set, the searched model and the hyper-parameter to obtain a trained prediction model.

Therefore, the model and the hyper-parameters are selected from the model search space, the model and the hyper-parameters which are more suitable for the target prediction scene can be preferentially selected, and the efficiency of determining the prediction model can be improved.

In one possible implementation, the search strategy further includes a training strategy; according to the target data set, the searched model and the hyper-parameter, model training is carried out to obtain a prediction model which is trained, and the method comprises the following steps: and carrying out model training according to the target data set, the training strategy, the searched model and the hyper-parameter so as to obtain a prediction model completing the training.

Therefore, the search strategy also comprises a training strategy, and the training strategy is used during training, so that the trained prediction model can be rapidly obtained, and the efficiency of determining the prediction model is further improved.

In a possible implementation manner, obtaining a search strategy corresponding to a target prediction scenario includes: determining a search strategy corresponding to a target prediction scene according to the target data set; or sending an acquisition request of the search strategy to the cloud device, wherein the acquisition request of the search strategy is used for requesting to acquire the search strategy; and receiving a search strategy sent by the cloud equipment.

According to the scheme, the construction equipment can determine the search strategy by itself and can request to acquire the search strategy from the cloud equipment, so that the search strategy can be flexibly acquired.

In a possible implementation manner, after obtaining the search strategy corresponding to the target prediction scenario, the method further includes: the search strategy is adjusted based on training experience in training the searched model. In this way, the trained predictive model can be obtained quickly, since the search strategy can be optimized.

In one possible implementation, the method further includes: and sending a prediction model corresponding to the target prediction scene, a model and a hyper-parameter used for training the prediction model corresponding to the target prediction scene, and an identification of the target prediction scene to the cloud device. Therefore, the construction equipment sends the relevant information of the prediction model corresponding to the target prediction scene to the cloud equipment, so that the cloud equipment can expand the model base and obtain more experiences of constructing the model.

In a possible implementation manner, before model training is performed according to a model and a hyper-parameter included in a target data set and a model search space to obtain a trained prediction model, the method further includes: acquiring a data preprocessing algorithm included in a model search space according to the target data set; obtaining a prediction model which finishes training according to a model and a hyper-parameter included in a target data set and a model search space, wherein the prediction model comprises: preprocessing a target data set according to a data preprocessing algorithm; and carrying out model training according to the model and the hyper-parameters included in the preprocessed target data set and the model search space so as to obtain a prediction model which completes training.

Therefore, the model search space also comprises a data preprocessing algorithm, so that the format of the data of the target data set can be matched with the model in the model search space, and the speed of obtaining the trained prediction model is accelerated.

In a second aspect, there is provided a method of constructing a predictive model, the method comprising: receiving a model acquisition request sent by construction equipment, wherein the model acquisition request is used for requesting to acquire an initial model search space corresponding to a target prediction scene; determining a model and a hyper-parameter corresponding to a target prediction scene according to a historical data set, wherein the historical data set is the target data set of the target prediction scene, or data in the historical data set is similar to data in the target data set; and sending an initial model search space to the construction equipment, wherein the initial model search space comprises a model corresponding to the target prediction scene and the hyper-parameters.

Therefore, the cloud device can provide an initial model search space for the construction device, the construction device does not need to be selected from a large number of models and hyper-parameters, and therefore the prediction model can be determined quickly.

In one possible implementation, determining a model and a hyper-parameter corresponding to a target prediction scenario includes: and obtaining a model and a hyper-parameter with the similarity between the characteristics of the applicable data set and the second characteristics of the target data set, and determining the model and the hyper-parameter in the initial model search space corresponding to the target prediction scene. In this way, an initial model search space corresponding to the target predicted scene may be obtained.

In a third aspect, the present application provides an apparatus for building a prediction model, where the apparatus includes a plurality of modules, and the modules implement the method for building a prediction model provided in the first aspect by executing instructions.

In a fourth aspect, the present application provides an apparatus for building a prediction model, which includes a plurality of modules, and the plurality of modules implement the method for building a prediction model provided in the second aspect by executing instructions.

In a fifth aspect, the present application provides a computing device comprising a memory and a processor, the processor executing computer instructions stored by the memory to cause the computing device to perform the method of building a predictive model of the first aspect described above.

In a sixth aspect, the present application provides a computing device comprising a memory and a processor, the processor executing computer instructions stored by the memory to cause the computing device to perform the method of building a predictive model of the second aspect described above.

In a seventh aspect, a computer-readable storage medium is provided, which stores computer instructions, and when the computer instructions in the computer-readable storage medium are executed by a computing device, the computing device is caused to execute the method for building a prediction model according to the first aspect, or the computing device is caused to implement the functions of the apparatus according to the third aspect.

In an eighth aspect, a computer-readable storage medium is provided, which stores computer instructions, and when the computer instructions in the computer-readable storage medium are executed by a computing device, the computing device is caused to execute the method for building a prediction model according to the second aspect, or the computing device is caused to realize the functions of the apparatus according to the fourth aspect.

In a ninth aspect, the present application provides a computer program product comprising computer instructions which, when executed by a computing device, perform the method of building a predictive model of the first aspect described above.

In a tenth aspect, the present application provides a computer program product comprising computer instructions which, when executed by a computing device, perform the method of building a predictive model of the second aspect described above.

In an eleventh aspect, the present application provides a system for building a prediction model, where the system includes a cloud device and a building device, where the building device is the apparatus according to the third aspect, and the cloud device is the apparatus according to the fourth aspect.

Drawings

FIG. 1 is a schematic block diagram of a computing device provided in an exemplary embodiment of the present application;

FIG. 2 is a schematic diagram of an application scenario for building a prediction model according to an exemplary embodiment of the present application;

FIG. 3 is a schematic diagram of an application scenario for building a prediction model according to an exemplary embodiment of the present application;

FIG. 4 is a schematic flow chart diagram of a method for constructing a predictive model provided by an exemplary embodiment of the present application;

fig. 5 is an interaction diagram of a cloud device and a building device when building a prediction model according to an exemplary embodiment of the present application;

FIG. 6 is a block diagram of a process for building a predictive model as provided by an exemplary embodiment of the present application;

FIG. 7 is a schematic structural diagram of an apparatus for constructing a predictive model according to an exemplary embodiment of the present application;

FIG. 8 is a schematic structural diagram of an apparatus for constructing a predictive model according to an exemplary embodiment of the present application;

fig. 9 is a schematic structural diagram of an apparatus for building a prediction model according to an exemplary embodiment of the present application.

Detailed Description

To make the objects, technical solutions and advantages of the present application more clear, embodiments of the present application will be described in further detail below with reference to the accompanying drawings.

To facilitate understanding of the embodiments of the present application, the following first introduces concepts of the terms referred to in the embodiments of the present application:

1. automated Machine Learning (AutoML) is a process of automating a predictive model that applies a Machine Learning model to real-world problems. In a typical machine learning application, the data is suitably processed to fit the data into a machine learning model, and then the machine learning model and the hyper-parameters are optimized using the data and the known machine learning model to obtain a final predictive model.

2. The time sequence is a sequence formed by sequencing numerical values of certain statistical indexes according to time sequence. The time series prediction method is to make and analyze time series, and analogize or extend according to the development process, direction and trend reflected by the time series, so as to predict the level which can be reached in the next period of time or in several years later. The time series prediction method comprises the following steps: the method comprises the steps of collecting and sorting historical data of a certain social phenomenon, checking and identifying the data to form a time series, analyzing the time series, searching for a rule that the social phenomenon changes along with time change to obtain a certain pattern, and then using the pattern to predict the future situation of the social phenomenon.

3. Hyper-parameters, in the context of machine learning, are parameters set prior to the start of learning, rather than parameter data obtained through training. Specifically, the hyper-parameter may be an initial value of a parameter in the model, or may be a parameter used for controlling a model training process in learning. For example, the hyper-parameters are a learning rate, an iteration step size, and the like.

With the development of network technology and the development of computing technology, the demands of operators and users for various content predictions continuously increase, and a large number of prediction models have been accumulated at present, however, there is a continuous prediction demand for different contents in different prediction scenes, so that it is necessary to provide a method for constructing prediction models to automatically select optimal models and hyper-parameters for different prediction durations, so as to realize the rapid generation of prediction models in various prediction scenes and improve the construction efficiency of prediction models. Based on the application, the method for constructing the prediction model is provided, and can be executed by a device for constructing the prediction model, which is hereinafter referred to as a construction device for short. The construction device may be a hardware device, such as a server, a terminal computing device, etc., or may be a software device (such as a set of software programs running on a hardware device). For example, the building apparatus may operate in a cloud computing device system (which may include at least one cloud computing device, such as a server, etc.), may also operate in an edge computing device system (which may include at least one edge computing device, such as a server, a desktop computer, etc.), and may also operate in various terminal computing devices (such as a notebook computer, a personal desktop computer, etc.).

The building means may logically be a means constituted by individual parts. The various components in the building apparatus may be deployed in different systems or servers, respectively. Each part of the construction device can be respectively operated in any two of the cloud computing equipment system, the edge computing equipment system and the terminal computing equipment. The cloud computing device system, the edge computing device system and the terminal computing device are connected through communication paths, and can communicate with each other and transmit data.

The embodiment of the application also provides computing equipment for constructing the prediction model. FIG. 1 illustratively provides one possible architecture diagram for a computing device 100. The computing device includes a memory 101, a processor 102, a transceiver 103, and a bus 104. The memory 101, the processor 102 and the transceiver 103 are connected to each other through a bus 104.

Memory 101 may be ROM, static storage, dynamic storage, or RAM. The memory 101 may store computer instructions that, when executed by the processor 102 stored in the memory 101, the processor 102 and the transceiver 103 are used to perform a method of building a predictive model. The memory may also store data, for example, a portion of memory 101 for storing data needed to build a predictive model, as well as for storing intermediate or result data during program execution.

The processor 102 may be a general purpose CPU, an application ASIC, a Graphics Processing Unit (GPU), or any combination thereof. The processor 102 may include one or more chips.

The transceiver 103 enables communication between the computing device and other devices or communication networks using transceiver modules such as, but not limited to, transceivers.

Bus 104 may include a path that transfers information between various components of the computing device (e.g., memory 101, processor 102, transceiver 103).

The embodiment of the application can be applied to various scenes, and two possible scenes are given as follows:

in a first scenario, as shown in fig. 2, the device for constructing a scenario of a prediction model includes a cloud device, a plurality of construction devices, and each construction device is used for constructing a prediction model of a different scenario. The cloud device is in communication connection with the construction device, and the construction device is in communication connection with the terminal device. The cloud computing device is used for determining a search space and a search strategy (the search strategy comprises a model search strategy and a training strategy) of a subsequently mentioned initial model and sending the search strategy to the construction device. The build device is a computing device such as a server. If the method for constructing the prediction model is applied to a network, the terminal device is a router, a switch, a base station and the like. The construction equipment is used for constructing a prediction model based on the initial model search space and the search strategy. The construction equipment is also used for sending the constructed prediction model to the terminal equipment. And deploying the constructed prediction model by the terminal equipment to perform prediction processing. In a scenario one, when the construction apparatus is a software program, the construction apparatus is deployed on the cloud device and the construction device. And when the construction device is a hardware device, the construction device is a cloud device and/or a construction device.

In a second scenario, as shown in fig. 3, the device for constructing a scenario of the prediction model includes a construction device and a terminal device, and the construction device is connected to the terminal device. The construction equipment is computing equipment such as a server, and if the method for constructing the prediction model is applied to a network, the terminal equipment is a router, a switch, a base station and the like. The construction device is used for storing historical data sets of various scenes, executing the process of constructing the prediction model, and then sending the constructed prediction model to the terminal device. And deploying the constructed prediction model by the terminal equipment to perform prediction processing. In scenario two, the subsequently mentioned initial model search space and search strategy are determined by the construction apparatus. In scenario two, when the construction apparatus is a software program, the software program is run on the construction apparatus. When the construction device is a hardware device, the construction device is a construction device.

In addition, in the embodiment of the present application, the initial model search space and the search strategy may also be determined by different devices. Specifically, the initial model search space is determined by the cloud device, and the search strategy is determined by the construction device, or the initial model search space is determined by the construction device, and the search strategy is determined by the cloud device.

In addition, in the embodiment of the application, the cloud device can also be used as a construction device to construct a prediction model, and the prediction model is sent to the terminal device and the like.

In the embodiment of the present application, the method for constructing the prediction model may be applied to the construction of prediction models of various contents, and the embodiment of the present application takes a time series in a prediction network as an example for explanation, that is, each piece of data in a data set mentioned later is a time series in a network. For example, the time series in the network may be a time series of packet loss rates, a time series of transmission delays, and the like. In addition, in the embodiments of the present application, "and/or" referred to hereinafter refers to three cases. For example, "A and/or B" refers to A, B, as well as A and B.

The method for constructing the prediction model provided by the embodiment of the present application will be described with reference to fig. 4. As shown in fig. 4, the processing flow of the method is as follows:

step 401, the construction device obtains a model search space corresponding to the target prediction scene according to the target data set of the target prediction scene and/or the scene information of the target prediction scene.

The target prediction scenario is any scenario, for example, the target prediction scenario is a prediction scenario of network traffic data, a prediction scenario of a Key Performance Indicator (KPI) of network packet loss (each KPI corresponds to one prediction scenario, or KPIs having common characteristics correspond to one prediction scenario), and the like. The target data set is a data set corresponding to a target prediction scene and comprises a large number of time series. The model search space includes a model and hyper-parameters associated with the target prediction scenario.

In this embodiment, the construction device may obtain a model search space corresponding to the target prediction scene by using the target data set of the target prediction scene. Or, the construction device may obtain a model search space corresponding to the target prediction scene by using scene information of the target prediction scene. Or, the construction device may obtain a model search space corresponding to the target prediction scene using the target data set and the scene information of the target prediction scene.

In a possible implementation manner, the construction device provides an interface for inputting scene information of the target prediction scene for a user, and the user inputs the scene information of the target prediction scene into the construction device through the interface, so that the construction device obtains the scene information of the target prediction scene. The context information of the target prediction context includes computational performance requirements and/or prediction requirements, the computational performance requirements including one or more of memory information, CPU information, or inference speed. Here, the memory information includes memory occupancy rate and/or memory occupancy amount, the memory occupancy rate refers to a ratio of memory occupied by the prediction model corresponding to the operation target prediction scene to the total memory of the deployed device, and the memory occupancy amount refers to a memory occupied by the prediction model corresponding to the operation target prediction scene. The CPU information comprises CPU occupancy rate and/or CPU occupancy amount, the CPU occupancy rate refers to the proportion of CPU resources occupied by the prediction model corresponding to the operation target prediction scene to the total CPU resources of the deployed equipment, and the CPU occupancy amount refers to the CPU resource amount occupied by the prediction model corresponding to the operation target prediction scene.

The prediction demand comprises prediction duration and/or prediction period which are predicted by using the constructed prediction model, the prediction duration refers to the duration of each prediction, and the prediction period refers to the prediction interval. For example, the prediction duration is 1 month, and the prediction period is 10 days, i.e., one prediction every 10 days, one month for each prediction.

Step 402, the construction equipment performs model training according to the model and the hyper-parameters included in the target data set and the model search space to obtain a prediction model after training.

In the present embodiment, at the start of training, the construction apparatus searches for the model and the hyper-parameters in the model search space in accordance with a model search strategy (described later). The initial objective function is stored in the construction equipment, and the objective function is used for constraining the training target of the prediction model. In addition, the build device provides an interface for the user to input training constraints for the predictive model, and the user can input additional training constraints. For example, the initial objective function is to constrain the prediction error of the prediction model, and the training constraint additionally input by the user may be a performance constraint (e.g., the prediction model processes a minimum of 50 time series in 1 second, the memory occupied when running the prediction model cannot exceed the target value, etc.). The construction equipment can combine the initial objective function and the training constraint additionally input by the user into a training target of the prediction model.

And when the construction equipment searches the model and the hyperparameter in the model search space each time, the training target is used as constraint, and model training is carried out according to the target data set of the target prediction scene and the model and the hyperparameter searched each time, so as to obtain a prediction model which completes training. And (3) each time the construction equipment obtains a prediction model, judging whether all combinations of the models and the hyper-parameters in the model search space are completely trained, if all combinations are completely trained, executing the step 403, and if not, returning to the step 402. When the prediction model is trained, a training strategy (described later) is used to determine whether parameter sharing is possible, when parameter sharing is possible, parameters of other models are shared, whether an early-stop feature (described later) occurs in the training process is determined, and if the early-stop feature occurs, the training is finished, and then the next model and the hyper-parameter combination are searched again to be trained, so that the most suitable prediction model is found quickly.

Here, in the process of training a certain model, if the model is a prediction model that has already been built, the training is to update parameters in the prediction model. If the model is a model framework (i.e., a machine learning algorithm), training is to determine the parameters in the model framework.

And 403, obtaining a prediction model corresponding to the target prediction scene by the construction equipment according to the evaluation result of the trained prediction model.

In this embodiment, the construction device evaluates each of the trained prediction models based on a preset evaluation mode to obtain an evaluation result of each prediction model. For example, the construction apparatus determines the amount of data processed per second for each prediction model that has completed training, as an evaluation result for each prediction model. And then the construction equipment selects a prediction model corresponding to the target prediction scene from the trained prediction models based on the evaluation result of the trained prediction models.

Optionally, in step 402 and step 403, the building apparatus may evaluate each time a trained prediction model is obtained. And target evaluation requirements of the prediction models are stored in the construction equipment, and when a certain prediction model is detected to meet the target evaluation requirements, the prediction model is determined as the prediction model corresponding to the target prediction scene. Therefore, the combination of all models and hyper-parameters in the model search space does not need to be trained, and the speed of determining the prediction model corresponding to the target prediction scene can be improved.

Therefore, the construction equipment can acquire the model search space of the target prediction scene, so that selectable models in the model search space are related to the target prediction scene, and compared with the models in the related technology, the number of models is small, so that all models and hyper-parameters in the model search space can be quickly combined to complete training, the construction time of the prediction model can be further saved, and the efficiency of determining the prediction model is improved.

The following will be described in addition to the flow shown in fig. 4:

in this embodiment, step 401 can be implemented in a variety of ways, and four possible ways are given as follows:

the first method is as follows: the construction equipment acquires an initial model search space from the cloud equipment, and determines a model search space corresponding to the target prediction scene according to the initial model search space.

In this embodiment, the building device sends a model obtaining request to the cloud device, where the model obtaining request is used to request to obtain the initial model search space, and the model obtaining request includes the target data set or a second feature of the target data set, where the second feature may be a partial feature or a full feature of the target data set. The cloud device stores a large number of models (the models may be already-constructed models or machine learning models (basic machine learning algorithms)) and hyper-parameters, and the corresponding already-constructed models store characteristics of data sets used for training the models, and the corresponding machine learning models store characteristics of applicable data sets. After the cloud device receives the model acquisition request, the cloud device determines an initial model search space using the target data set or the second feature of the target data set (the determination process is described later). And the cloud device sends the initial model search space to the construction device.

And the construction equipment receives the initial model search space sent by the cloud equipment. The construction equipment screens out the model and the hyper-parameters in the initial model search space by using the scene information of the target prediction scene and/or the first characteristics of the target data set, and forms a model search space corresponding to the target prediction scene.

And secondly, the construction equipment determines an initial model search space, and determines a model search space corresponding to the target prediction scene according to the initial model search space.

In this embodiment, a large number of models (which may be already-constructed models or machine learning models) and hyper-parameters are stored in the construction device, and the features of the data set used for training the models and the applicable scenario information are stored in correspondence with the already-constructed models, and the features of the applicable data set and the applicable scenario information are stored in correspondence with the machine learning models. The construction equipment determines an initial search space (the determination mode is the same as that of the cloud-end equipment in the first mode), and then the model and the hyper-parameters are screened in the initial model search space by using the scene information of the target prediction scene and/or the first characteristics of the target data set to form a model search space corresponding to the target prediction scene.

Optionally, in the first and second manners, the process of determining the initial search space is:

the cloud device is used as an example to describe here, the cloud device determines the second characteristic of the target data set, and selects, from a large number of models, a model in which the second characteristic of the data set is similar to the second characteristic of the target data set. Specifically, the second feature of the data set used by the constructed model (which refers to the model that does not belong to the target prediction scene and has been trained) in the screened model during training is similar to the second feature of the target data set, and the second feature of the data set applicable to the machine learning model in the screened model is similar to the second feature of the target data set. And the cloud equipment determines the hyper-parameters corresponding to the screened models, and forms an initial model search space by using the screened models and the hyper-parameters.

Or the cloud device may construct a plurality of models by using the target data set and the plurality of machine learning models and the hyper-parameters, and select the machine learning models and the hyper-parameters (N may be 30, etc.) corresponding to the first N models with the best evaluation results from the plurality of models to form an initial model search space.

In the above first and second modes, the process of the construction device for screening the model and/or the hyper-parameter in the initial model search space is as follows:

the construction equipment selects a model and a hyper-parameter from the initial model search space to form a model search space corresponding to the target prediction scene. Specifically, the construction device screens a model with a first feature of the applicable data set similar to a first feature of the target data set in the initial model search space, and determines a hyper-parameter corresponding to the model (i.e., screens a hyper-parameter with a first feature of the applicable data set similar to a first feature of the target data set). And the construction equipment forms a model search space corresponding to the target prediction scene by using the screened model and the hyper-parameters corresponding to the model. Specifically, the first feature of the data set used by the constructed model in the screened model during training is similar to the first feature of the target data set, and the first feature of the data set applicable to the machine learning model in the screened model is similar to the first feature of the target data set.

Or, the construction equipment screens a model with scene information similar to the scene information of the target prediction scene in the initial model search space, and determines the hyper-parameters corresponding to the model (i.e. screens out the hyper-parameters with applicable scene information similar to the target prediction scene). And the construction equipment forms a model search space corresponding to the target prediction scene by using the screened model and the hyper-parameters corresponding to the model.

Or, in the initial model search space, the construction device screens a model in which the scene information has similarity with the scene information of the target prediction scene and the first feature of the applicable data set has similarity with the first feature of the target data set, and determines a hyper-parameter corresponding to the model (i.e., screens out a hyper-parameter in which the applicable scene information is similar to the target prediction scene and the first feature of the applicable data set is similar to the first feature of the target data set). And the construction equipment forms a model search space corresponding to the target prediction scene by using the screened model and the hyper-parameters corresponding to the model.

It should be noted here that the first characteristic of the target data set may be the same as or different from the second characteristic of the target data set. Specifically, when the construction device performs the screening process on the model and/or the hyper-parameter in the initial model search space only by using the scene information of the target prediction scene, the first feature of the target data set and the second feature of the target data set may be the same, and may be all the features or partial features of the target data set. When the construction equipment performs screening processing on the model and/or the hyper-parameters in the initial model search space by using at least the first characteristic of the target data set, the first characteristic of the target data set is different from the second characteristic of the target data set. For example, the first feature of the target data set is classification information and/or statistical information mentioned later. The second characteristic of the target data set is the maximum, minimum, and average values of the data in the data set, etc.

It should be further noted here that, when the model and/or the hyper-parameter in the initial model search space is/are screened, only the model in the initial model search space may be screened, the screened model is taken as the model in the model search space, and the hyper-parameter in the initial model search space is directly taken as the hyper-parameter in the model search space. Of course, it may also be that only the hyper-parameters in the initial model search space are screened, the screened hyper-parameters are used as the hyper-parameters in the model search space, and the model in the initial model search space is directly used as the model in the model search space.

In a third mode, the construction equipment can directly use the scene information and/or the target data set of the target prediction scene to determine the model search space corresponding to the target prediction scene.

In this embodiment, a large number of models (which may be already-constructed models or machine learning models) and hyper-parameters are stored in the construction device, and the features of the data set used for training the models and the applicable scenario information are stored in correspondence with the already-constructed models, and the features of the applicable data set and the applicable scenario information are stored in correspondence with the machine learning models. The construction equipment uses the scene information of the target prediction scene, and screens out a model with scene information similar to the scene information of the target prediction scene from a large number of models. And the construction equipment determines the hyper-parameters corresponding to the screened model, and the screened model and the hyper-parameters form a model search space. Or, the construction device uses the features of the target data set (which may be a combination of the first feature and the second feature), selects a model with similarity between the features of the applicable data set and the features of the target data set from a large number of models, determines the hyper-parameters corresponding to the selected model, and forms a model search space with the selected model and the hyper-parameters. Or, the construction equipment uses the characteristics of the target data set and the scene information of the target prediction scene, and screens out a model in which the scene information is similar to the scene information of the target prediction scene and the characteristics of the applicable data set are similar to the characteristics of the target data set from a large number of models. And the construction equipment determines the hyper-parameters corresponding to the screened model, and the screened model and the hyper-parameters form a model search space.

And fourthly, the construction equipment requests the model search space corresponding to the target prediction scene from the cloud equipment. The cloud device may directly use the scene information and/or the target data set of the target prediction scene to determine a model search space corresponding to the target prediction scene (see manner three for the determination). And the cloud equipment sends the model search space to the construction equipment.

The first characteristic of the target data set may include classification information and/or statistical information. The classification information refers to cycle information, fluctuation information, mutation information, and the like of each time series in the target data set. The statistical information includes the sampling interval, sampling duration, and miss-acquire condition (i.e., non-acquire condition) for each time series of the target data set. For the period information, it can be classified as having periodicity and not having periodicity, that is, for each time series, the time series has periodicity or not. When the time series has periodicity, the period information may further include a period length.

In one possible implementation manner, in step 403, the construction apparatus determines evaluation results of a plurality of trained prediction models, and then determines a prediction model with an optimal evaluation result among the plurality of prediction models. And the construction equipment determines the prediction model as a prediction model corresponding to the target prediction scene. Specifically, when the evaluation result is determined, the accuracy of the model and the performance of the model are weighted to obtain a weighted value, and the prediction model with the largest weighted value is determined as the prediction model with the optimal evaluation result.

In a possible implementation manner, in this embodiment of the application, before step 402, the building device may obtain the model search strategy and the training strategy in the following manner:

the construction equipment obtains a search strategy corresponding to the target prediction scene, wherein the search strategy comprises a model search strategy and/or a training strategy.

In this embodiment, the construction device may obtain a search strategy corresponding to the target prediction scenario by using the target data set and/or the scenario information of the target prediction scenario, where the search strategy includes a model search strategy and/or a training strategy. The search strategy is used to indicate how quickly and accurately to find the optimal predictive model.

Specifically, the model search strategy includes a model and hyper-parameter selection strategy for selecting a model and a hyper-parameter in the model search space. The training strategy is used to increase the training speed during the training process. The training strategy may include hyper-parameters (such as iteration step length), early-stop features (such as when a certain model is used for training, after a certain early-stop feature occurs, the model is stopped from being trained, that is, the model is considered not to be used for training to obtain a prediction model), parameter sharing, low fidelity, and the like. The early-stop feature may be that the number of iterations reaches a target number, the update amount of the parameter is below a certain threshold, and so on. Parameter sharing refers to sharing parameters among different models during training. Low fidelity may refer to training using a portion of the data of the target data set, increasing training speed, and the like.

In a possible implementation manner, in this embodiment of the present application, the process of constructing the device to obtain the search policy includes:

the construction equipment determines a search strategy corresponding to a target prediction scene according to the target data set; or the construction equipment sends a search strategy acquisition request to the cloud equipment, wherein the search strategy acquisition request is used for requesting to acquire a search strategy; and receiving a search strategy sent by the cloud equipment.

In this embodiment, a large number of model search strategies and training strategies are stored in the construction device, and the characteristics of the applicable data sets are stored corresponding to each model search strategy and each training strategy respectively. The construction equipment can determine a model search strategy and a training strategy corresponding to the characteristics of the target data set according to the target data set.

Or, the construction device sends an acquisition request of the search policy to the cloud device, where the acquisition request includes the target data set or the characteristics of the target data set. A large number of search strategies are stored in the cloud device, and the characteristics of the applicable data set are stored corresponding to each search strategy. The cloud device determines a model search strategy and a training strategy (the determination mode is the same as that of the construction device), and sends the model search strategy and the training strategy to the construction device.

It should be noted here that the request for obtaining the search policy and the request for obtaining the model mentioned above may be the same request, or may not be the same request, and the embodiment of the present application is not limited. The features of the target data set may be the classification information and/or statistical information (i.e., first features) mentioned above. Of course, the second feature mentioned above is also possible. Of course, a union of the first feature and the second feature is also possible.

In addition, the cloud device issues the model search strategy and the training strategy to the construction device at the same time, and certainly, the model search strategy and the training strategy can also be issued respectively.

The corresponding construction equipment obtains a search strategy, the construction equipment uses a model search strategy corresponding to the target prediction scene in the processing of step 402, the processing of optimally selecting the model and the hyper-parameters in the model search space can be realized, the most suitable model and the hyper-parameters are preferentially selected, and in the training process, the training performance is improved through the training strategy (for example, training is finished on some models in advance), so the training speed can be further improved.

In a possible implementation manner, the construction device may further adjust the search policy, and the process is as follows:

and adjusting the search strategy according to the training experience when the searched model is trained.

In this embodiment, when the building apparatus trains the model in the training search space, the search strategy is adjusted based on the training experience of the training. For example, when some models in the training search space are trained, the evaluation results of the models are relatively poor, and some feature (i.e., early-stop feature) appears in the training process of the models, and the feature is added to the training strategy. And subsequently, in the training process of other models, when the feature appears again, the training can be directly finished. And subsequently, using the adjusted search strategy in the process of training the model. For another example, when a model in the training search space is trained, the evaluation result of the model is poor, the type of the model is added to the model search strategy, and then a model different from the type of the model is preferentially selected in the model search space. In this way, the training speed can be increased since the search strategy can be adjusted.

In one possible implementation, after step 403 (building the prediction model), the building device sends the prediction model corresponding to the target prediction scenario, the model and the hyper-parameter used for training the prediction model corresponding to the target prediction scenario, and the identifier of the target prediction scenario to the cloud device.

In this embodiment, after the building device builds the prediction model, the building device sends the built prediction model, the model and the hyper-parameter used for training the prediction model, and the identifier of the target prediction scenario to the cloud device (e.g., a cloud storage device). And after receiving the built prediction model, the model and the hyper-parameter used for training the prediction model and the identification of the target prediction scene, the cloud device correspondingly stores the built prediction model, the model and the hyper-parameter used for training the prediction model and the identification of the target prediction scene. Therefore, the trained prediction model can be conveniently expanded by the cloud equipment, and experience is provided for construction of other models.

In addition, the construction equipment can also provide the performance of the constructed prediction model, the accuracy of the prediction model and the prediction duration to the cloud equipment. For example, the performance of the predictive model is such that the time taken to process 1000 pieces of data is 1 second, etc. The accuracy of the prediction model is measured using a variety of metrics. For example, accuracy is measured using index 1 and index 2, where index 1 is Mean Square Error (MSE), index 2 is percentage absolute Error, and so on. The MSE is equal to the mean of the squares of the differences between the predicted and actual values over the predicted duration. The percentage absolute error is equal to the ratio of the absolute value of the difference between the actual value and the predicted value to the actual value. The prediction duration is used to indicate the length of time that the prediction model can predict.

In addition, the build device may also provide the cloud device with the target dataset (or characteristics of the target dataset) and the search strategy used to build the prediction model. For example, the data may be characterized as having periodicity.

As shown in table one, the content provided by the building device to the cloud device is provided, including model identification of the prediction model, model name of the prediction model, accuracy of the prediction model, characteristics of a target data set used for building the prediction model, prediction duration, performance of the prediction model, and the like:

watch 1

In a possible implementation manner, after step 403 (building a prediction model), the building device sends a model deployment message to the terminal device, where the model deployment message is used to indicate that the network device that receives the model deployment message deploys the prediction model.

In this embodiment, after the construction device constructs the prediction model, it generates a model deployment message, and the model deployment message carries the constructed prediction model. And then the construction equipment sends the model deployment message to the terminal equipment. And the terminal equipment receives the model deployment message, analyzes the model deployment message to obtain a constructed prediction model, and stores the prediction model. And the subsequent terminal equipment predicts the content to be predicted by using the prediction model.

Optionally, when the building device sends the model deployment message to the terminal device, the building device may also send the predicted duration and/or the predicted period. Specifically, the model deployment message may be carried in the model deployment message. If the predicted duration and/or the predicted period are not included in the model deployment message, the terminal device may use a default predicted duration and/or predicted period.

Optionally, before inputting the data into the prediction model, when the data may need to be preprocessed, the construction device may further send a data preprocessing algorithm to the terminal device. When the subsequent terminal equipment uses the prediction model, the data preprocessing algorithm is firstly used for preprocessing the data, and then the preprocessed data are input into the prediction model.

Optionally, when the terminal device detects data using the prediction model, the terminal device may send the data to the construction device, so that the construction device expands the target data set. In order to enable the cloud device to obtain a large amount of data, the construction device may send the data received from the terminal device to the cloud device.

In a possible implementation manner, in the model search space, the format of the data required to be input by each model and the like may be different, or the different formats of the data may affect the training result, so that the model search space may further include a data preprocessing algorithm. For example, the data preprocessing algorithm is an interpolation complement algorithm, a data smoothing algorithm, and the like. In this way, the initial model search space provided by the cloud device also includes a data preprocessing algorithm. Thus, in step 401, the model search space obtained by the construction equipment further includes a data preprocessing algorithm, and the corresponding initial model search space also includes the data preprocessing algorithm.

Specifically, when a data preprocessing algorithm in the initial model search space is determined, a plurality of data preprocessing algorithms, a plurality of machine learning models and a plurality of hyper-parameters are stored in the cloud device. The cloud device obtains a target data set and determines a combination of various data preprocessing algorithms, a machine learning model and hyper-parameters. And the cloud equipment performs model training to obtain a plurality of models by using each combination and the target data set. And then the cloud equipment evaluates the multiple models to obtain a target number of models with the optimal evaluation result, and a combination of a data preprocessing algorithm, a machine learning model and a hyper-parameter used for obtaining the target number of models is added to the initial model search space.

And/or the cloud device analyzes the target data set to obtain the characteristics of the target data set. The cloud device then uses the features to determine a constructed model that yields a similarity of the features of the applicable dataset to the features of the target dataset. The cloud device then determines the data preprocessing algorithm used in obtaining the constructed model. And the cloud equipment adds the constructed model, the data preprocessing algorithm and the hyper-parameters corresponding to the constructed model to an initial model search space. The characteristics of the target data set are here the same as in the previous paragraph.

In this way, the initial model search space sent by the cloud device to the construction device includes the data preprocessing algorithm. When the construction equipment constructs a prediction model corresponding to a target prediction scene, firstly, a data preprocessing algorithm is used for preprocessing a target data set, and then model training is carried out by using the preprocessed target data set, a searched model and a hyper-parameter, so that a prediction model which is trained is obtained.

The cloud device determines the data preprocessing algorithm, but may also determine the data preprocessing algorithm by the construction device in the same manner as the cloud device.

In addition, the construction device obtains the initial model search space and the search strategy from the cloud device, as shown in fig. 5, an interaction process between the cloud device and the construction device is also provided:

step 501, the construction device sends a model obtaining request to the cloud device (taking the model obtaining request and the obtaining request of the search policy as the same request as an example), where the model obtaining request includes the target data set or the second feature of the target data set.

Step 502, the cloud device determines an initial model search space and a search strategy corresponding to the target prediction scene based on the target data set or the second characteristic of the target data set.

Step 503, the cloud device sends the initial model search space and the search strategy corresponding to the target prediction scene to the construction device.

Step 504, the construction equipment receives an initial model search space and a search strategy corresponding to the target prediction scene.

And 505, screening the model and/or the hyper-parameter in the initial model search space by the construction equipment based on the target data set and/or the scene information of the target prediction scene to obtain a model search space corresponding to the target prediction scene.

Step 506, the construction equipment obtains a prediction model corresponding to the target prediction scene based on the search strategy, the training target of the prediction model, the target data set and the model search space.

Step 507, the construction device sends a prediction model corresponding to the target prediction scene, an identifier of the target prediction scene, and the like to the cloud device.

In the application, the model corresponding to the target prediction scene is obtained, not all machine learning models, so that when the prediction model of the target prediction scene is trained, the number of models needing to be trained is small, processing resources can be saved, and the efficiency of constructing the prediction model can be improved. And because the model in the model search space is a model related to the target prediction scene, a prediction model with higher performance is easier to acquire.

In addition, for better understanding of the embodiments of the present application, the embodiments of the present application further provide a frame diagram for constructing a prediction model, as shown in fig. 6:

assuming that the target data set needs to be preprocessed, data preprocessing is performed on data in the target data set, and a model search space and a search strategy are obtained based on the target data set and/or scene information of a target prediction scene. The construction device then determines a training objective for the predictive model. The construction equipment trains a prediction model based on a training target, a search strategy and a model search space. And the construction equipment evaluates the prediction model to obtain the prediction model corresponding to the target prediction scene. In addition, the construction equipment can update the search strategy based on the evaluation result so as to improve the speed of constructing the prediction model.

In addition, the scenario described in the embodiment of the present application is a prediction scenario, and may also be a scenario for detecting an anomaly, so that the prediction model mentioned in the foregoing is a prediction model for detecting an anomaly.

Fig. 7 is a schematic structural diagram of building a prediction model according to an embodiment of the present application. The apparatus may be implemented as part or all of an apparatus in software, hardware, or a combination of both. The apparatus provided in the embodiment of the present application may implement the process described in fig. 4 in the embodiment of the present application, and the apparatus includes: an acquisition module 710, a training module 720, and a determination module 730, wherein:

an obtaining module 710, configured to obtain a model search space corresponding to a target prediction scene according to a target data set of the target prediction scene and/or scene information of the target prediction scene, where the model search space includes a model and a hyper-parameter, and may be specifically used to implement the obtaining function of step 401 and an implicit step included in step 401;

a training module 720, configured to perform model training according to the model and the hyper-parameters included in the target data set and the model search space to obtain a trained prediction model, which may be specifically used to implement the training function of step 402 and the implicit step included in step 402;

the determining module 730 is configured to obtain the prediction model corresponding to the target prediction scenario according to the evaluation result of the trained prediction model, and may specifically be configured to implement the determining function in step 403 and the implicit step included in step 403.

In a possible implementation manner, the obtaining module 710 is configured to:

obtaining an initial model search space, wherein the initial model search space is obtained by analyzing a historical data set, the historical data set is a target data set of the target prediction scene, or data in the historical data set is similar to data in the target data set; screening models and/or hyper-parameters in the initial model search space according to the target data set and/or the scene information of the target prediction scene to obtain a model search space corresponding to the target prediction scene, wherein the models in the model search space are partial models or all models in the initial model search space, and the hyper-parameters in the model search space are partial hyper-parameters or all hyper-parameters in the initial model search space; alternatively, the first and second electrodes may be,

and obtaining a model search space corresponding to the target prediction scene in a preset model and a preset hyper-parameter according to a target data set of the target prediction scene and/or scene information of the target prediction scene.

In one possible implementation form of the method,

the model in the model search space is a model with similarity between the characteristics of the data set applicable in the initial model search space and the first characteristics of the target data set, and the hyper-parameters in the model search space are hyper-parameters with similarity between the characteristics of the data set applicable in the initial model search space and the first characteristics of the target data set; alternatively, the first and second electrodes may be,

the model in the model search space is a model with similarity between the scene information applicable to the initial model search space and the scene information of the target prediction scene, and the hyper-parameters in the model search space are hyper-parameters with similarity between the scene information applicable to the initial model search space and the scene information of the target prediction scene; alternatively, the first and second electrodes may be,

the model in the model search space is a model in which the scene information applicable in the initial model search space has similarity with the scene information of the target prediction scene and the characteristics of the applicable data set have similarity with the first characteristics of the target data set, and the hyper-parameters in the model search space are hyper-parameters in which the scene information applicable in the initial model search space has similarity with the scene information of the target prediction scene and the characteristics of the applicable data set have similarity with the first characteristics of the target data set.

In a possible implementation manner, the obtaining module 710 is configured to:

determining a model and a hyper-parameter corresponding to the target prediction scene according to the historical data set, and forming an initial model search space by the determined model and the hyper-parameter; alternatively, the first and second electrodes may be,

sending a model obtaining request to cloud equipment, wherein the model obtaining request is used for requesting to obtain the initial model searching space; and receiving an initial model search space sent by the cloud equipment.

In a possible implementation manner, the obtaining module 710 is configured to:

and obtaining a model and a hyper-parameter with the similarity between the characteristics of the applicable data set and the second characteristics of the target data set, and determining the model and the hyper-parameter in the initial model search space corresponding to the target prediction scene.

In one possible implementation, the second characteristic of the target data set is the same characteristic or a different characteristic than the first characteristic of the target data set.

In a possible implementation manner, the context information of the target prediction scenario includes a calculation performance requirement and/or a prediction requirement, the calculation performance requirement includes one or more of memory information, CPU information, or inference speed, and the prediction requirement includes a prediction duration and/or a prediction period of prediction performed by a prediction model corresponding to the target prediction scenario.

In one possible implementation, the first feature of the target data set includes classification information and/or statistical information, wherein the classification information includes one or more of period information, fluctuation information or mutation information of a time series in the target data set, and the statistical information includes one or more of sampling interval, sampling duration or missing acquisition condition of the time series in the target data set.

In a possible implementation manner, the determining module 730 is configured to:

and selecting a prediction model with the optimal evaluation result from the trained prediction models, and determining the prediction model as the prediction model corresponding to the target prediction scene.

In a possible implementation manner, the obtaining module 710 is further configured to:

obtaining a search strategy corresponding to the target prediction scene before obtaining a trained prediction model according to the model and the hyperparameter which are included in the target data set and the model search space, wherein the search strategy comprises a model search strategy;

the training module 720 is configured to:

searching a model and a hyper-parameter in the model search space according to the model search strategy;

and carrying out model training according to the target data set and the searched model and the hyper-parameter so as to obtain a prediction model which is trained.

In one possible implementation, the search strategy further includes a training strategy;

the training module 720 is configured to:

and carrying out model training according to the target data set, the training strategy, the searched model and the hyper-parameter so as to obtain a prediction model completing the training.

In a possible implementation manner, the obtaining module 710 is configured to:

determining a search strategy corresponding to the target prediction scene according to the target data set; alternatively, the first and second electrodes may be,

the device further comprises: a sending module 740, configured to send an acquisition request of a search policy to a cloud device, where the acquisition request of the search policy is used to request to acquire the search policy; and the receiving module is used for receiving the searching strategy sent by the cloud equipment.

and after a search strategy corresponding to the target prediction scene is obtained, adjusting the search strategy according to training experience for training a searched model.

In one possible implementation, as shown in fig. 8, the apparatus further includes:

the sending module 740 is configured to send, to the cloud device, the prediction model corresponding to the target prediction scenario, the model and the hyper-parameter used for training the prediction model corresponding to the target prediction scenario, and the identifier of the target prediction scenario.

acquiring a data preprocessing algorithm included in the model search space according to the target data set before acquiring a trained prediction model according to the model and the hyperparameter included in the target data set and the model search space;

the training module 720 is configured to:

preprocessing the target data set according to the data preprocessing algorithm;

and carrying out model training according to the preprocessed target data set and the model and the hyper-parameters included in the model search space to obtain a prediction model which completes training.

The division of the modules in the embodiments of the present application is illustrative, and only one logical function division is provided, and in actual implementation, there may be another division manner, and in addition, each functional module in each embodiment of the present application may be integrated in one processor, may also exist alone physically, or two or more modules are integrated into one module. The integrated module can be realized in a hardware mode, and can also be realized in a software functional module mode.

Fig. 9 is a schematic structural diagram of building a prediction model according to an embodiment of the present application. The apparatus may be implemented as part or all of an apparatus in software, hardware, or a combination of both. The device includes: a receiving module 910, a determining module 920 and a sending module 930, wherein:

a receiving module 910, configured to receive a model obtaining request sent by a construction device, where the model obtaining request is used to request to obtain an initial model search space corresponding to a target prediction scene, and specifically may be used to implement a receiving function of a device for constructing a prediction model;

a determining module 920, configured to determine a model and a hyper-parameter corresponding to the target prediction scenario according to a historical data set, where the historical data set is a target data set of the target prediction scenario, or data in the historical data set is similar to data in the target data set, and specifically may be used to implement a determining function of a device for constructing a prediction model;

a sending module 930, configured to send an initial model search space to the building apparatus, where the initial model search space includes a model and a hyper-parameter corresponding to the target prediction scenario, and may be specifically configured to implement a sending function of a device for building a prediction model.

In a possible implementation manner, the determining module 920 is configured to:

In the above embodiments, all or part of the implementation may be realized by software, hardware, firmware or any combination thereof, and when the implementation is realized by software, all or part of the implementation may be realized in the form of a computer program product. The computer program product comprises one or more computer program instructions which, when loaded and executed on a server or terminal, cause the processes or functions described in accordance with embodiments of the application to be performed, in whole or in part. The computer instructions may be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another, for example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wire (e.g., coaxial cable, fiber optics, digital subscriber line) or wirelessly (e.g., infrared, wireless, microwave, etc.). The computer readable storage medium can be any available medium that can be accessed by a server or a terminal or a data storage device, such as a server, a data center, etc., that incorporates one or more of the available media. The usable medium may be a magnetic medium (such as a floppy Disk, a hard Disk, a magnetic tape, etc.), an optical medium (such as a Digital Video Disk (DVD), etc.), or a semiconductor medium (such as a solid state Disk, etc.).

Claims

1. A method of constructing a predictive model, the method comprising:

acquiring a model search space corresponding to a target prediction scene according to a target data set of the target prediction scene and/or scene information of the target prediction scene, wherein the model search space comprises a model and a hyper-parameter;

performing model training according to the model and the hyper-parameters included in the target data set and the model search space to obtain a prediction model after training;

and obtaining a prediction model corresponding to the target prediction scene according to the evaluation result of the trained prediction model.

2. The method according to claim 1, wherein the obtaining a model search space corresponding to a target prediction scene according to a target data set of the target prediction scene and/or scene information of the target prediction scene comprises:

3. The method of claim 2, wherein the model in the model search space is a model in which features of the dataset applicable in the initial model search space have similarity with first features of the target dataset, and the hyper-parameters in the model search space are hyper-parameters in which features of the dataset applicable in the initial model search space have similarity with first features of the target dataset; alternatively, the first and second electrodes may be,

4. The method of claim 2 or 3, wherein the obtaining an initial model search space comprises:

5. The method of claim 4, wherein determining the model and the hyper-parameters corresponding to the target prediction scenario from the historical data set comprises:

6. The method of claim 5, wherein the second characteristic of the target dataset is the same characteristic or a different characteristic than the first characteristic of the target dataset.

7. The method according to any one of claims 1 to 6, wherein the scene information of the target prediction scene comprises a calculation performance requirement and/or a prediction requirement, the calculation performance requirement comprises one or more of memory information, Central Processing Unit (CPU) information or inference speed, and the prediction requirement comprises a prediction duration and/or a prediction period for prediction by a prediction model corresponding to the target prediction scene.

8. The method of any one of claims 2 to 7, wherein the first characteristic of the target data set comprises classification information and/or statistical information, wherein the classification information comprises one or more of period information, fluctuation information or mutation information of a time series in the target data set, and the statistical information comprises one or more of sampling interval, sampling duration or missing acquisition condition of the time series in the target data set.

9. The method according to any one of claims 1 to 8, wherein obtaining the prediction model corresponding to the target prediction scenario according to the evaluation result of the trained prediction model comprises:

10. The method according to any one of claims 1 to 9, wherein before performing model training based on the model and the hyper-parameters included in the target data set and the model search space to obtain the trained predictive model, the method further comprises:

obtaining a search strategy corresponding to the target prediction scene, wherein the search strategy comprises a model search strategy;

the obtaining of the trained prediction model according to the model and the hyper-parameters included in the target data set and the model search space includes:

searching a model and/or a hyper-parameter in the model search space according to the model search strategy;

and carrying out model training according to the target data set, the searched model and the hyper-parameter so as to obtain a prediction model which is trained.

11. The method of claim 10, wherein the search strategy further comprises a training strategy;

the model training according to the target data set, the searched model and the hyper-parameter to obtain the trained prediction model comprises the following steps:

12. The method according to claim 10 or 11, wherein the obtaining of the search strategy corresponding to the target prediction scenario includes:

sending an acquisition request of a search strategy to cloud equipment, wherein the acquisition request of the search strategy is used for requesting to acquire the search strategy; and receiving the search strategy sent by the cloud equipment.

13. The method according to any one of claims 10 to 12, wherein after the obtaining of the search strategy corresponding to the target prediction scenario, the method further comprises:

the search strategy is adjusted based on training experience in training the searched model.

14. The method according to any one of claims 1 to 13, further comprising:

and sending a prediction model corresponding to the target prediction scene, a model and a hyper-parameter used for training the prediction model corresponding to the target prediction scene, and an identifier of the target prediction scene to cloud equipment.

15. The method according to any one of claims 1 to 14, wherein before performing model training based on the model and the hyper-parameters included in the target data set and the model search space to obtain the trained predictive model, the method further comprises:

acquiring a data preprocessing algorithm included in the model search space according to the target data set;

the model training according to the model and the hyper-parameters included in the target data set and the model search space to obtain a trained prediction model comprises:

16. A method of constructing a predictive model, the method comprising:

receiving a model obtaining request sent by construction equipment, wherein the model obtaining request is used for requesting to obtain an initial model search space corresponding to a target prediction scene;

determining a model and a hyper-parameter corresponding to the target prediction scene according to a historical data set, wherein the historical data set is the target data set of the target prediction scene, or data in the historical data set is similar to data in the target data set;

and sending an initial model search space to the construction equipment, wherein the initial model search space comprises a model and a hyper-parameter corresponding to the target prediction scene.

17. The method of claim 16, wherein determining the model and the hyper-parameters corresponding to the target prediction scenario from the historical data set comprises:

18. An apparatus for constructing a predictive model, the apparatus comprising:

the system comprises an acquisition module, a storage module and a processing module, wherein the acquisition module is used for acquiring a model search space corresponding to a target prediction scene according to a target data set of the target prediction scene and/or scene information of the target prediction scene, and the model search space comprises a model and a hyper-parameter;

the training module is used for carrying out model training according to the model and the hyper-parameters included in the target data set and the model search space so as to obtain a prediction model which is trained;

and the determining module is used for obtaining the prediction model corresponding to the target prediction scene according to the evaluation result of the trained prediction model.

19. The apparatus of claim 18, wherein the obtaining module is configured to:

20. The apparatus of claim 19, wherein the model in the model search space is a model in which a feature of the dataset applicable in the initial model search space has similarity to a first feature of the target dataset, and the hyper-parameter in the model search space is a hyper-parameter in which a feature of the dataset applicable in the initial model search space has similarity to a first feature of the target dataset; alternatively, the first and second electrodes may be,

21. The apparatus of claim 19 or 20, wherein the obtaining module is configured to:

22. The apparatus of claim 21, wherein the obtaining module is configured to:

23. The apparatus of claim 22, wherein the second characteristic of the target dataset is the same characteristic or a different characteristic than the first characteristic of the target dataset.

24. The apparatus according to any one of claims 18 to 23, wherein the context information of the target prediction context includes a computational performance requirement and/or a prediction requirement, the computational performance requirement includes one or more of memory information, Central Processing Unit (CPU) information or inference speed, and the prediction requirement includes a prediction duration and/or a prediction period for prediction by a prediction model corresponding to the target prediction context.

25. The apparatus of any one of claims 20 to 24, wherein the first characteristic of the target data set comprises classification information and/or statistical information, wherein the classification information comprises one or more of period information, fluctuation information or mutation information of a time series in the target data set, and the statistical information comprises one or more of sampling interval, sampling duration or missing acquisition condition of the time series in the target data set.

26. The apparatus of any one of claims 18 to 25, wherein the determining module is configured to:

27. The apparatus according to any one of claims 18 to 26, wherein the obtaining module is further configured to:

performing model training according to the model and the hyper-parameters included in the target data set and the model search space to obtain a search strategy corresponding to the target prediction scene before obtaining a prediction model after training, wherein the search strategy comprises a model search strategy;

the training module is configured to:

28. The apparatus of claim 27, wherein the search strategy further comprises a training strategy;

the training module is configured to:

29. The apparatus of claim 27 or 28, wherein the obtaining module is configured to:

30. The apparatus according to any one of claims 27 to 29, wherein the obtaining module is further configured to:

31. The apparatus of any one of claims 18 to 30, further comprising:

and the sending module is used for sending the prediction model corresponding to the target prediction scene, the model and the hyper-parameter used for training the prediction model corresponding to the target prediction scene, and the identification of the target prediction scene to the cloud equipment.

32. The apparatus according to any one of claims 18 to 31, wherein the obtaining module is further configured to:

the training module is configured to:

33. An apparatus for constructing a predictive model, the apparatus comprising:

the model acquisition module is used for requesting to acquire an initial model search space corresponding to a target prediction scene;

the determining module is used for determining a model and a hyper-parameter corresponding to the target prediction scene according to a historical data set, wherein the historical data set is the target data set of the target prediction scene, or data in the historical data set is similar to data in the target data set;

and the sending module is used for sending an initial model search space to the construction equipment, wherein the initial model search space comprises a model corresponding to the target prediction scene and a hyper-parameter.

34. The apparatus of claim 33, wherein the determining module is configured to:

35. A computing device to build a predictive model, the computing device comprising a processor and a memory, wherein:

the memory having stored therein computer instructions;

the processor executes the computer instructions to cause the computing device to perform the method of any of claims 1 to 17.

36. A computer-readable storage medium storing computer instructions which, when executed by a computing device, cause the computing device to perform the method of any of claims 1 to 15 or to implement the functionality of the apparatus of any of claims 18 to 32.

37. A computer-readable storage medium storing computer instructions which, when executed by a computing device, cause the computing device to perform the method of claim 16 or 17 or to implement the functionality of the apparatus of claim 33 or 34.